1. Field of the Invention
This invention generally relates to systems and methods for inspecting and measuring structures created during the fabrication of semiconductor devices.
2. Description of the Related Art
The following description and examples are not admitted to be prior art by virtue of their inclusion in this section.
The semiconductor industry has been using optical critical dimension (CD) metrology (such as scatterometry) since about 2000, but many of the current uses in high-volume manufacturing are limited to measurement of relatively simple shapes, usually just a grid of parallel trenches or structures, and measurement of relatively few shape parameters such as height (depth), CD (width) and sidewall angle.
In order to make a measurement, a model of the structure has to be constructed. Usually cross-section electron micrographs of the structure are needed because, in most cases, the shapes cannot be determined from top-down images. If the shape of the structure is three-dimensional (i.e. the structure does not have a constant cross-section any direction), then at least two perpendicular cross-sections may be needed to reveal the shape.
A model is most often constructed from simple geometric shapes that approximate the shape of the structure. The dimensions of these shapes are controlled by a few parameters (such as length, width, height and/or angles). When setting up the model, a decision has to be made as to which of these dimensional parameters will be allowed to vary during the measurement process and which will be kept constant.
Values or models of the complex refractive indices of the materials that make up the structure are needed. In many cases, these will be known from prior experience with these materials or by measurements at unpatterned locations on the same wafer or from other wafers processed through the same, or similar, equipment and processes.
Once the shapes, dimensions and refractive indices are known, electromagnetic calculations can predict how light will scatter from that structure. Those scattering predictions can be used to model the expected signal when an optical instrument makes a measurement of that structure.
The complete model (sometimes referred as a measurement recipe) is then used to process data collected on an optical measuring tool such as a reflectometer or ellipsometer in order to determine the best fitting shape parameters, which are assumed to represent the relevant dimensions of the actual shape.
In many cases, the model may be used to precompute a library of optical signatures corresponding to ranges of all the dimensional parameters that are allowed to vary. A library may speed up the measurement significantly when more than 2 or 3 parameters are allowed to vary.
It is also known to construct libraries of optical signatures from experimentally measured optical signatures collected by measuring structures on wafers that were processed under different conditions. In some cases, other measurement techniques, such as SEM images, are used also to determine some of the dimensions.
The need for cross-section images means that an accurate model cannot be constructed for many hours or even days after the first wafers have been processed because of the time needed to prepare the wafers for cross-sectioning as well as the time required for taking the images. This delay is generally not acceptable, and the cost is high. Often initial measurements have to be made using models constructed before cross-section images are available and so those models incorporate a lot of guesswork and may in not provide accurate measurements for the structure. If the results subsequently prove to be accurate, until the cross-section images become available, there may be a lack of confidence in the results leading to delays in acting upon those results.
Two perpendicular cross-section images plus a top-down image may not suffice to reveal all the details of complex structures made from multiple materials. Re-entrant features, in particular, may be missed unless a cross-section happens to go through the right location.
Since cross-sections are slow and expensive to prepare, typically only a few will be prepared. These will not show all the possible variations in shapes and dimensions that can occur with normal variations in processing, let alone the changes that may occur when abnormal situations arise.
As described above, the dimensions of the geometric shapes that make up the model are controlled by a set of parameters (such as length(s), width(s), height(s) and/or angles). When setting up the recipe, decisions have to be made which dimensional parameters should be kept fixed and which should be allowed to vary during the measurement process. If many, or all, parameters are allowed to vary in an attempt to maximize the flexibility of the model to track process changes, the measurement results will usually exhibit poor repeatability (and for 20 or more parameters may be unstable) because the optical signal may poorly discriminate between certain combinations of dimensional changes. But if one or more parameters are held constant when the corresponding dimensions are actually varying, then the measurement results will be inaccurate.
The process of constructing the model of the structure involves a combination of experience, guesswork and trial and error and is, at best, a stow process that is not consistent from person to person, and, at worst, may not result in an accurate measurement.
When a library is constructed from experimental data, the library cannot be constructed until multiple wafers have been fully processed under different process conditions and the optical measurements have been performed on those wafers. Such a library suffers from the disadvantages of being noisy. Firstly, there is process noise because, even for the same process settings on the process tool, the actual processing conditions do vary with location on the wafer and from wafer to wafer. Secondly, there is necessarily noise on the optical measurements from optical, thermal and electrical noise sources in the instrument. Thirdly, any reference dimensional or shape measurements (from, for example an electron micrograph or an atomic force microscope) are also subject to noise and systematic errors.
Accordingly, it would be advantageous to develop process aware metrology systems and/or methods that do not have one or more of the disadvantages described above.